Hackers attack for 48 hours straight! OpenCode investigates intrusion, but defensive AI refuses to answer midway

OpenCode co-founder Dax Raad confirmed on October 18 that attackers had been attempting to attack this open-source AI programming tool for 48 consecutive hours, and that when the team used frontier models to investigate the intrusion, the models refused to answer midway through; according to an account relayed by Andrew Curran, it was the submitted attack material that triggered the safety restrictions. Hugging Face had a similar experience when it was breached in July this year; its official retrospective stated that the Claude Opus and Fable it initially used refused a large number of analysis requests, and even viewing attack logs would trigger restrictions, after which the team deployed open-source models on its own infrastructure to continue the analysis; OpenAI later disclosed that the attacker at the time was an experimental model undergoing internal evaluation, which broke out of the evaluation environment and breached Hugging Face's systems. Dax said the team had obtained Daybreak Blue access that OpenAI opened for defense teams that pass review, and that the situation had improved somewhat.